LLM-augmented multimodal radar reflectivity retrieval with cross-satellite transfer learning

Multimodal radar reflectivity retrieval plays a crucial role in enabling robust extreme-weather monitoring and cross-satellite observational integration under heterogeneous meteorological conditions. Existing retrieval methods typically rely on deep learning to learn end-to-end mappings from satellite observations, or apply domain adaptation to reduce distribution gaps. However, these approaches often rely excessively on low-level statistical alignment. They fail to explicitly characterize differences in observation configurations and physical semantics across satellites, which results in performance degradation in cross-satellite transfer scenarios. Moreover, redundant information and background noise in satellite data further limit feature extraction and generalization under complex weather conditions. To address these issues, this paper proposes MUSTNet, a multi-source, multimodal radar reflectivity retrieval framework. The proposed framework integrates residual encoding, spatial–channel attention, and multi-scale feature interaction into a unified multimodal retrieval network. It incorporates LLM-based semantic priors to establish a semantically conditioned gating mechanism, embedding high-level observational and meteorological semantics into the architecture. In addition, a cross-satellite consistency-driven transfer learning framework is developed. Through staged source-domain pretraining and progressive target-domain fine-tuning, it enables high-quality fusion and structurally consistent reconstruction between geostationary satellite and ground-based radar data. Extensive experiments on real-world datasets illustrate superior performance compared with the state-of-the-art approaches.